NVIDIA Now Allows GeForce GPU Pass-Through For Windows VMs On Linux

[oiaohm]

You don't have to have 2 GPUs. In fact, I find the whole concept with 2 GPUs ridiculous, you need 2 GPUs, two displays (or KVM switch, maybe you can use different inputs on the display, depending on the monitor used ofc.), multiple inputs, it really only makes sense for software development, where you want to do it in the VM while having host working. If you take the most common usage of VMs passthrough (Windows guest for gaming), it really makes 0 sense to use 2 GPUs or to even have host "active" (in the sense of DE, processes etc.), the ideal situation in such scenario would be to detach the primary GPU, attach it to the VM, kill host session (free up resources) and reverse the process when done with VM (gaming).

https://looking-glass.io/

The idea that you need multi displays or KVM switch is part wrong. It is possible with the higher end cards enterprise/pro grade cards to pass the output from KVM instance back to a Window on the host instance.

Ideal processes have the important word ideal in them. The reality is you normally never want a ideal process because it normally has no allowance if something goes wrong.

Host having a screen is kind of important if the VM does something really stupid like causing the GPU to over heat to the point you cannot switch back to it..
2 GPU do make sense really to deal with things have not gone as expected. Remember lots of modern laptops have 2 GPU one connected to the screen being intel or AMD and Nvidia gpu working with PRIME support passing that output across. If the host has maintained the intel/AMD intergrated GPU control and the VM has the decrete GPU...

There is a problem here Nvidia and others does not what output less cards like those sold to mining being in the market for desktop users wanting a second GPU for VMs. There is really no reason why GPU for a VM has to have outputs other than vendor decided limitations and some software to make PRIME support between GPUs are more usable tool.

[RBilettess]

I hope AMD follows this decision. I tried to setup "PCI passthrough via OVMF" with my 5700XT card and the AMD windows driver gave me "error 43", no matter what workaround I have tried :-(

[Buntolo]

It was extremely easy to work around the code 43 error. All you had to do was stick a

[FONT=monospace][COLOR=#000000]<kvm> [/COLOR]
<hidden state='on'/>
</kvm>[/FONT]

in your KVM config.

I've been trying for months reading tons of guides:
1) that didn't work for me
2) it's not that easy

[intelfx]

What's the point of the passthrough then? Wasn't it possible already to dedicate one GPU to a virtual machine?

That's exactly what passthrough is about.

[kdemello1980]

I've been trying for months reading tons of guides:
1) that didn't work for me
2) it's not that easy

There are other issues, mostly related to iommu groups and PCIE ACS, which are CPU related, rather than GPU related. If you have enough PCIE root ports, and a properly wired motherboard, these are non-issues. If moving the to-be-passed-through GPU to another slot doesn't work, then sometimes the only fix for these is new hardware.

At one time Xen was more forgiving of ACS issues, and it worked for me, but I'm not sure if it is anymore.

[ThoreauHD]

You don't need 2 gpus.

https://looking-glass.io/

And nvidia just reverted their driver back to what it was before they introduced 'error 43' to sell more quadros while attempting to audit datacenters.

[dragon321]

I hope AMD follows this decision. I tried to setup "PCI passthrough via OVMF" with my 5700XT card and the AMD windows driver gave me "error 43", no matter what workaround I have tried :-(

On the other hand my RX 570 worked fine with Windows on VM.

[phoronix_is_awesome]

There is a problem here Nvidia and others does not what output less cards like those sold to mining being in the market for desktop users wanting a second GPU for VMs. There is really no reason why GPU for a VM has to have outputs other than vendor decided limitations and some software to make PRIME support between GPUs are more usable tool.

Nvidia surely does not want outputless cards like mining cards to be used as a desktop card for VM. But arguably, it is a better use for mining cards to do graphics and CUDA compute instead of stupid-ass-mining.

[RBilettess]

On the other hand my RX 570 worked fine with Windows on VM.

So AMD should continue to officially block their drivers in a VM, because it worked for you? Or what do you want to say?

[zexelon]

I've been trying for months reading tons of guides:
1) that didn't work for me
2) it's not that easy

There are a myriad of variables. Most likely related to your motherboard setup. The error 43 issue with Nvidia is very easy to bypass... it is really nice to know now though that it may not require bypassing in the near future!

I have run GPU passthrough on an i5 for many years... used the integrated HD4600 for linux and passed through the 1070. It was totally solid and made so many things nicer to manage wrt windows. The only issue I ran into was some more modern games (i.e. warzone and star craft 2 ... so basically activision/blizzard) detected the hypervisor as a cheat system and either soft banned (warzone) or would not run at all (i.e. SC2).

I now have an threadripper system so unfortunately the convenience of embedded GPU is no longer there and that is annoying, I now will need 2x GPUs or a sepperate device (i.e. laptop and ssh) to configure and get it going but either way its something I am hoping to go at again.